Regional multi-user load prediction method for coupling neural network and time sequence model

Abstract

The invention discloses a regional multi-user load prediction method for coupling a neural network and a time sequence model. The method comprises the steps of data acquisition and preprocessing, multi-user feature construction, shared feature extraction, individual feature learning, feature fusion output and coupling model parameter rolling updating. The core lies in that cross-user sharing features are extracted through a multi-channel neural network, individual differences are captured in combination with a pre-training time sequence model, multi-user prediction results are output in parallel after fusion, and rapid parallel prediction of multi-user loads in a regional range is achieved. The method can significantly reduce training overhead, improve prediction precision and response speed, and provide technical support for demand response scheduling and low-carbon operation optimization in a novel power system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction and control technology of power systems, specifically involving a regional multi-user load prediction method that couples neural networks and time series models. It is applicable to load prediction and demand response optimization scenarios of multi-building and distributed users in new power systems. Background Technology

[0002] With the large-scale integration of renewable energy, energy storage systems, and electric vehicles, power systems are gradually exhibiting characteristics of high volatility, high uncertainty, and strong time-varying nature. The operating status of regional energy systems changes in real time with factors such as weather, user behavior, and distributed power output. In order to maintain supply and demand balance and improve energy utilization efficiency, load forecasting needs to be updated rapidly to support real-time grid dispatching and demand response decisions.

[0003] Against this backdrop, rapid load forecasting for multiple users across regions has become a crucial step. However, existing load forecasting methods are mostly designed for single buildings / users, resulting in long forecasting cycles and high computational costs, making them unsuitable for complex scenarios involving simultaneous forecasting of multiple users across a region. Common load forecasting methods, such as traditional algorithms based on statistical laws (e.g., linear regression, ARIMA models) and neural network methods based on deep learning (e.g., LSTM, CNN, GRU), can achieve relatively ideal forecasting accuracy in single-user scenarios, but often struggle to balance accuracy and efficiency in multi-user or regional load forecasting. The main reason is that single-user modeling methods require building and training models separately for each user, leading to enormous computational overhead and complex model maintenance when the number of users is large. While clustering modeling and transfer learning methods can reduce the number of models to some extent, they can easily weaken the expression of individual features or introduce negative transfer phenomena, affecting forecasting stability and generalization ability.

[0004] In recent years, with the rapid development of large-scale models, time-series models (such as Informer, TimeGPT, and Timer) have demonstrated excellent performance in capturing long-term dependencies and recognizing complex sequence patterns. However, the application of these models in multi-user load forecasting still has limitations. They lack information-sharing mechanisms among different users, and in multi-user scenarios, they usually require individual fine-tuning, resulting in high computational costs and difficulty in capturing common regional features. Consequently, it is difficult to balance prediction efficiency and generalization performance in large-scale scenarios.

[0005] Therefore, there is an urgent need to propose a fast load forecasting method for regional multi-user / building applications with shared learning and parallel forecasting capabilities, so as to balance forecasting accuracy, speed and model scalability, thereby meeting the actual needs of regional multi-user load forecasting and demand response control in smart grids. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a regional multi-user load forecasting method that couples neural networks and time-series models. This method introduces a shared feature extraction layer based on a multi-channel convolutional neural network into the model architecture to capture the common patterns of multiple users, and establishes a lightweight individual feature modeling layer through a time-series model to preserve user differences. This enables fast and high-precision parallel forecasting of multi-user loads within a region, and can support the real-time scheduling and optimization needs of new power systems.

[0007] A method for predicting regional multi-user load by coupling a neural network and a time-series model includes the following steps: A method for predicting regional multi-user load using a coupled neural network and a time-series model, characterized by the following steps: Step 1: Collect historical load data and environmental parameters from multiple users in the area, and perform data preprocessing to ensure data integrity and availability; Step 2: Based on the established prediction target, construct a multi-user feature set, which includes shared features and individual features; Step 3: Extract shared features across users using a multi-channel neural network; Step 4: Capture individual characteristics of each user using a time-series model; Step 5 involves fusing shared features and individual features and then inputting the fused features into the fully connected layer of the coupled model to output the prediction results. Step 6: Train and optimize the parameters of the coupled model based on the loss function; Step 7 involves deploying the optimized coupling model in the regional energy system for real-time rolling forecasting.

[0008] Step 1 involves collecting historical load data and environmental parameters from multiple users in the area and performing data preprocessing to ensure data integrity and availability. Specifically, this involves collecting historical load data and environmental parameters from multiple users in the area. Environmental parameters include hourly weather data and corresponding time labels. Data preprocessing includes handling outliers and missing values. Specifically, outlier location is based on the distribution of load data at the same time each day. Data deviating from the distribution range is identified as outliers and replaced using missing value imputation. Missing value imputation uses linear interpolation, specifically by using the arithmetic mean of historical data values ​​from the same time of the same week and the data value from the time preceding the time to be imputed as the interpolation result.

[0009] The specific method for extracting cross-user shared features through a multi-channel convolutional neural network in step 3 is as follows: a shared feature extraction layer is constructed based on a multi-channel convolutional neural network. The shared feature extraction layer achieves cross-user feature association learning by sharing convolutional kernel parameters in the user dimension, and introduces a temporal channel reweighting and normalization mechanism so that the convolutional neural network can maintain the alignment and comparability of load patterns between different users while expanding the temporal receptive domain. The output of the shared feature extraction layer is a fused region-level shared feature vector, which is used for subsequent fusion with the output features of individual feature modeling.

[0010] The specific method for capturing individual characteristics of each user through the time series model in step 4 is as follows: The time series model adopts different parameter freezing strategies according to computing resources and application requirements: When training and deployment are required, all parameters of the time series model are frozen to achieve coupled model training and deployment; when prediction accuracy needs to be improved, some parameters of the time series model are unfrozen to improve accuracy while achieving lightweight adaptation.

[0011] The process of establishing the coupling model in step 5 specifically includes: Structurally, the shared feature extraction layer based on a multi-channel neural network and the individual feature modeling layer of the time series model are integrated end-to-end. The shared feature vector output by the multi-channel neural network is flattened and then concatenated and fused with the individual feature embedding vectors of each user in the feature dimension to form a fused feature input. Finally, the fused feature is mapped by a fully connected layer and then connected to the output neurons to achieve parallel prediction of the load of each user. Step 6, which trains and optimizes the parameters of the coupled model based on the loss function, specifically includes: The coupled model is trained end-to-end. A suitable error loss function is selected or designed for the training objective function. The parameters of the coupled model are updated during the backpropagation phase. The parameters of the shared layer are shared and updated synchronously among all users to capture common regional features. The unfrozen parameters of the individual layer are updated independently according to the historical load characteristics of each user. Training is terminated when the coupled model reaches the convergence criterion. The convergence criterion is that the decrease in the loss function of the coupled model on the validation set is less than a preset threshold within a preset number of consecutive training cycles.

[0012] The loss function in step 6 is either the error loss function used for training conventional prediction models, or a hybrid loss function composed of global prediction error and individual error selected to improve training results. The weight coefficients of the hybrid loss function are adaptively adjusted according to the fluctuation characteristics of user load.

[0013] The coupled model has a prediction step size of 1 to 96 time steps, supporting short-term day-ahead and intraday parallel load forecasting for multiple users in the region.

[0014] In summary, the technical solutions conceived by this invention have the following main advantages over the prior art: 1. The method of this invention achieves multi-user parallel prediction by coupling traditional multi-channel convolutional neural networks with time series models. It adopts an end-to-end integrated structure to realize integrated training and optimization from input load data to prediction results, avoiding the high computational overhead of training the model separately for each user, and achieving high prediction efficiency.

[0015] 2. The method proposed in this invention uses a shared feature layer to extract common regional patterns, while using a personalized layer to retain user-specific features. This ensures the prediction accuracy of the model under different user types and time periods, and demonstrates strong generalization performance. 3. The end-to-end regional multi-user load forecasting architecture adopted in this invention facilitates model deployment, supports rapid access and online fine-tuning for new users, and is user-friendly and convenient for large-scale regional power operation and scheduling tasks. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0017] Figure 2 This invention relates to a model architecture diagram consisting of a shared feature layer and a personalized feature layer. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions involved in this invention will be clearly and completely described below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. It should be understood that the specific embodiments described herein are only used to explain this invention and are not intended to limit this invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0019] This invention provides a regional multi-user load forecasting method that couples a neural network with a time series model. The specific implementation steps of this method are as follows: Figure 1 As shown, the specific embodiments can be described as follows: Step 1: Collect historical load data and environmental parameters of multiple users in the area, and perform data preprocessing to ensure data integrity and availability; In a specific embodiment of the present invention, real-time power consumption data of 30 building users in the area are collected through a smart grid terminal, with a sampling period of 1 hour, and weather data and corresponding time tags (including dry-bulb temperature, wet-bulb temperature, etc.) for the same period are downloaded from a meteorological website. Outlier handling: Based on the distribution of load data at the same time every day, the specific embodiment of the present invention combines the 3σ principle to locate outliers, and identifies data points that deviate from the mean by more than three standard deviations as outliers; Missing value imputation: For missing or outlier data, imputation is performed based on linear interpolation. Specifically, the arithmetic mean of the load values ​​of the same week type and time period in historical data and the load value of the time period to be imputed is taken as the interpolation result.

[0020] Step 2: Based on the established prediction target, construct a multi-user feature set, which includes shared features and individual features. Multi-user feature set construction: The prediction target is determined to be the load of each user 24 hours before the current day. That is, the output of the coupled model is the load of each of the 30 users in the next 24 time steps, and the total output dimension is 24×30. The input feature set of the coupled model is constructed, which includes shared feature input and individual feature input: the shared feature input of the coupled model is the hourly load sequence of 30 users in the past 3 days, stacked according to the user dimension to form a three-dimensional vector with dimensions of 3×24×30; the individual feature input of the coupled model is the single sequence data composed of each user's own historical load and external variables; the entire sample set after processing is divided into training set, validation set and test set in a 7:1:2 ratio.

[0021] Step 3: Extract cross-user shared features using a multi-channel neural network. Specifically, a shared feature extraction layer is constructed based on a multi-channel convolutional neural network. This layer achieves cross-user feature association learning by sharing convolutional kernel parameters across user dimensions. A temporal channel reweighting and normalization mechanism is introduced to ensure the convolutional neural network maintains alignment and comparability of load patterns between different users while expanding the temporal receptive domain. The output of the shared feature extraction layer is a fused region-level shared feature vector, used for subsequent fusion with individual feature modeling output features. The shared feature layer module designed in this example includes two dilated convolutional layers to receive three-dimensional shared input features, with 100 and 200 channels respectively. Batch normalization is added to improve training stability and speed. The output of this module is a fused region-level shared feature vector. Step 4: Capture individual characteristics of each user through a time series model. Specifically, the time series model employs different parameter freezing strategies based on computing resources and application requirements: when training and deployment are needed, all parameters of the time series model are frozen to achieve coupled model training and deployment; when prediction accuracy needs to be improved, some parameters of the time series model are unfrozen to improve accuracy while achieving lightweight adaptation. This example designs a personalized feature modeling layer that accepts single-sequence input features from each user for user-specific feature learning. Specifically, a Timer model is used to process the historical time series of each user, and a fully connected layer receives external environmental variables. The parameters of the Timer model are frozen to reduce model training time. Step 5: The fused shared features and individual features are input into the fully connected layer of the coupled model to output the prediction result. The coupled model is structurally built upon an end-to-end integration of a shared feature extraction layer constructed from a multi-channel neural network and an individual feature modeling layer from a time-series model. This achieves integrated modeling and optimization from input load data to prediction results. Specifically, the shared feature vector output by the multi-channel neural network is flattened and then concatenated with the individual feature embedding vectors of each user along the feature dimension to form a fused feature input. This fused feature is merged through a fully connected layer, and the fusion content includes the shared features, the Timer model, and the influence of external environmental variables transmitted by the fully connected layer. Finally, the output neurons are connected to achieve parallel load prediction for each user. In this embodiment, the system can output the load prediction results for 30 users for the next 24 hours in parallel, thereby supporting regional-level multi-user short-term load prediction tasks. Specifically: Feature fusion: After flattening the shared feature vector output in step 3, it is concatenated and fused with the individual feature embedding vectors of each user output in step 4 along the feature dimension to form a fused feature. Predicted output: The fused features are input into the fully connected layer of the coupled model, and finally connected to the output neuron to output the load prediction values ​​for 30 users in the next 24 hours in parallel. Step 6: Train and optimize the coupled model based on the loss function: Perform end-to-end training on the coupled model. A suitable error loss function is selected or designed as the training objective function. The coupled model parameters are updated during the backpropagation phase. Shared layer parameters are shared and updated synchronously among all users to capture common regional features; unfrozen parameters in the individual layers are updated independently based on each user's historical load characteristics. Training terminates when the coupled model reaches a convergence criterion, which is: within a preset number of consecutive training cycles, the decrease in the loss function of the coupled model on the validation set is less than a preset threshold. Specifically: For the coupled model training configuration, the Adam optimizer is used, with an initial learning rate of 1e-4 and a batch size of 16. The loss function consists of two parts: global mean squared error and individual weighted error, with the weight coefficients adaptively adjusted according to user load fluctuations. The entire coupled model is trained end-to-end. During backpropagation, the parameters of the shared feature layer are shared and updated synchronously among all users; the parameters of the individual feature layer are frozen and not updated. The convergence criterion set during training is to trigger an early stopping mechanism and terminate training when the model's loss function decreases less than a preset threshold for multiple consecutive training epochs on the validation set. In this example, the maximum training epoch is set to 500 epochs.

[0022] Step 7: Deploy the optimized coupled model in the regional energy system for real-time rolling prediction. According to the test, the training time of this coupled model is 13.6 seconds, which is about 15 times faster than the traditional single-user model, 75 times faster than the fine-tuning speed of the per-user large model, and about 18% lower in average absolute error compared with the traditional ensemble model.

[0023] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0024] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A regional multi-user load forecasting method coupling a neural network with a time series model, characterized in that, The method comprises the following steps: Step 1: Collecting historical load data and environmental parameters of multiple users in a region and performing data preprocessing to ensure data completeness and availability; Step 2: Based on the established prediction target, a multi-user feature set is constructed, including shared features and individual features; Step 3: Extracting shared features across users through a multi-channel neural network; Step 4: Capturing individual features of each user through a time series model; Step 5: After fusing the shared features and individual features, input the fused features into the fully connected layer of the coupling model to output the prediction results; Step 6: Training and parameter optimization of the coupling model based on the loss function; Step 7: Deploying the optimized coupling model in the regional energy system for real-time rolling prediction.

2. The regional multi-user load forecasting method of coupling a neural network with a time series model according to claim 1, wherein, In step 1, historical load data and environmental parameters of multiple users in a region are collected and data preprocessing is performed to ensure data completeness and availability. The specific method is as follows: historical load data and environmental parameters of multiple users in a region are collected, and the environmental parameters specifically include hourly weather data and corresponding time labels. Data preprocessing includes outlier and missing value processing. Specifically, outlier positioning is based on the distribution of load data at the same time of the day to identify data deviating from the distribution interval as outliers and replace them with missing value filling method. Linear interpolation method is used for missing value filling, which uses the arithmetic mean of the data value at the same time of the same week and the data value at the previous time of the time to be filled as the interpolation result. 3.The regional multi-user load forecasting method of coupling neural network with time series model according to claim 2, wherein, In step 3, the specific method of extracting shared features across users through a multi-channel convolutional neural network is as follows: a shared feature extraction layer is constructed based on a multi-channel convolutional neural network. The shared feature extraction layer shares convolution kernel parameters in the user dimension to realize cross-user feature association learning, and introduces time channel reweighting and normalization mechanism to enable the convolutional neural network to expand the time series perception domain while maintaining the alignment and comparability of load patterns between different users. The output of the shared feature extraction layer is the fused regional-level shared feature vector, which is used for subsequent feature fusion with individual feature modeling.

4. The regional multi-user load forecasting method of coupling a neural network with a time series model according to claim 3, wherein, In step 4, the specific method of capturing individual features of each user through a time series model is as follows: the time series model uses different parameter freezing strategies according to computing resources and application requirements. When training and deployment are needed, all parameters of the time series model are frozen to realize coupling model training and deployment. When the prediction accuracy needs to be improved, some parameters of the time series model are unfrozen to improve the accuracy while realizing lightweight adaptation.

5. The regional multi-user load forecasting method of coupling a neural network with a time series model according to claim 4, wherein, The establishment process of the coupling model in step 5 specifically includes: The shared feature extraction layer constructed based on the multi-channel neural network and the individual feature modeling layer of the time series model are integrated end-to-end; wherein the shared feature vector output by the multi-channel neural network is flattened and concatenated with the individual feature embedding vector of each user in the feature dimension to form a fused feature input; finally, the fused feature is mapped through a fully connected layer and connected to an output neuron to realize parallel prediction of the load of each user.

6. The regional multi-user load forecasting method of coupling a neural network with a time series model according to claim 5, wherein, In step 6, the coupling model is trained and parameter optimized based on the loss function, specifically including: The end-to-end training is performed on the coupling model, a proper error loss function is selected or designed as the training target function, and the coupling model parameters are updated in the back propagation stage, wherein the shared layer parameters are shared and synchronously updated among all users to capture regional common features, the individual layer parameters are independently updated according to the historical load characteristics of each user, and the training is terminated when the coupling model reaches a convergence standard, wherein the convergence standard is that, in a preset plurality of continuous training periods, the loss function of the coupling model on the verification set decreases by less than a preset threshold.

7. The regional multi-user load forecasting method of coupling a neural network with a time series model according to claim 6, wherein, The loss function of step 6 is an error loss function selected for conventional prediction model training, or a hybrid loss function composed of global prediction error and individual error selected to improve training effect, and the weight coefficient of the hybrid loss function is adaptively adjusted according to the load fluctuation characteristics of the user.

8. The regional multi-user load forecasting method of coupling a neural network with a time series model according to claim 1, wherein, The coupling model prediction step length is 1 to 96 time steps, supporting regional multi-user short-term day-ahead and intra-day parallel load prediction.

Images